Checks the op kernels, the concat fast path, the VRAM arbiter incl. every
shutdown path that has caused a production incident, and prefill/decode
against a reference build. Prints PASS/FAIL per check.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
GGML_VK_CONCAT_FASTPATH=0 forces the generic path and =log reports each
fast-path dispatch, so validation can prove the path fires at prefill and
not at decode.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Two ways a shutting-down server kept the VRAM arbiter's token, stalling every
other model:
- SIGTERM with a request in flight: terminate() stopped only the task queue.
The HTTP handler kept waiting for results the loop would never send, the main
thread blocked joining it, and the token was released only in the destructor
after that join - i.e. never, until the client went away. Terminate the
result queue too (readers now get "server is shutting down": a 503, or an SSE
error event), and let wait_until_no_sleep give up once stopping.
- SIGTERM while waiting for the token: the loop thread sat in a blocking flock
inside the wake callback, holding mutex_tasks, so terminate() waited for the
current holder to finish and then for a pointless restore of our weights.
Poll the lock non-blocking and abort the wait on shutdown; the aborted wake
stays cold, skips the restore, and the loop exits before update_slots.
Also release the token right after start_loop() returns rather than in the
destructor, and only after the device buffers are actually freed: the old
order unlocked before destroy(), so the next model could start uploading into
VRAM that was still occupied. Freeing at exit uses a new
llama_context_discard_device(), which drops weights, KV, MoE cache and scratch
without capturing host shadows first.
Off production: SIGTERM mid-generation or mid-wait now exits in ~0.2 s with
exit 0 and frees the token within 5-70 ms (previously never / after the
holder's whole generation). Handoff latency and greedy output are unchanged.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A SIGTERM that lands while the start_loop() thread is inside a sleeping
callback deadlocks the server. The callbacks run under
server_queue::mutex_tasks, and the kernel delivers the signal to that same
thread; the handler calls terminate(), which locks mutex_tasks again, and the
thread waits on itself forever.
With the VRAM arbiter that window is wide - a whole device release, or a
blocking wait for the GPU token - and llama-swap sends SIGTERM to evict
models, so it hit production: an evicted model wedged mid-release, never
released token.lock, and every other model queued behind it. SIGTERM could
not stop it either; only SIGKILL did.
Make the handler only write the signal number to a self-pipe, which is
async-signal-safe, and run shutdown_handler on a dedicated thread. terminate()
there simply waits until the loop thread drops mutex_tasks, after which the
loop exits normally. The thread is joined before the objects the handler uses
are destroyed. If pipe() fails the old direct call is kept.
Reproduced 6/6 off production (SIGTERM during release and during the token
wait); with the fix 8/8 clean exits, and handoffs, image requests and greedy
output are unchanged.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Upstream #28334 removed --mmap/--no-mmap/--mlock and llama-bench's -mmp, so the
harness flags no longer parse: the server would refuse to start and llama-bench
would silently produce no rows. Switch both harnesses to -lm none, the same
load mode, and note --moe-cache-mib as a server-only option worth measuring.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PZz44SLQvTXMyWGio6t9DZ
The MoE cache (--moe-cache-mib, #29887) keeps the most recently used host
experts in device banks. The on-demand VRAM path (release_device /
restore_device, used by the server VRAM arbiter and the on-demand mmproj path)
did not know about it, so its banks were never freed on handoff. Restoring them
by reallocation alone would also be wrong: the banks come back zeroed while the
slot maps still mark experts as cached, so the model would silently read zero
weights.
Free the banks on release and empty the cache instead of shadowing it: re-init
every LRU, re-point its slot-map pointers at the host slot maps and clear those
to -1. The banks are reallocated (which also re-inits their views) and cleared
on restore, before the scheduler is rebuilt, and the cache refills from the
host experts as before. Release runs in both modes, weights-only and evict_kv.
device_resident() now covers the cache too, for the early returns in
release/restore, the auto-restore in decode() and
llama_context_weights_resident().
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PZz44SLQvTXMyWGio6t9DZ
release_device(evict_kv) drops the scheduler, so a context evicted by the VRAM
arbiter and then destroyed reaches the destructor with a null sched. The upstream
"compute buffer size matches expectation" loop calls
ggml_backend_sched_get_buffer_size() unconditionally, whose GGML_ASSERT(sched)
then aborts: the server exits 134 instead of 0 on every shutdown taken while
cold, which under llama-swap makes an ordinary stop look like a crashed child and
leaves a ggml backtrace in the log each time. It also skipped the rest of the
destructor, so the cold teardown path had never actually run to completion.
Guard at the call site rather than relaxing the assert - the assert is right, and
every other caller reserves the scheduler first. Same shape as the null guards in
synchronize(), memory_breakdown() and the released-buffer iteration.
The loop is a diagnostic size comparison, already reported by sched_reserve() at
load, so skipping it when cold loses nothing.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PZz44SLQvTXMyWGio6t9DZ
The doorbell was edge-triggered: a waiter wrote <arena>/doorbell/<pid> and the
holder's warden turned the inotify event into a queue flag. A ring carried no
notion of "still wanted", so it was lost whenever the holder was not already warm
and listening - during its model load (the watch does not exist yet, and the
kernel does not queue events for a watch that is not there) or inside
restore_device - and any ring that survived past the point its sender had been
satisfied caused a spurious release. Lost rings wedged the waiter permanently,
because flock(LOCK_EX) never times out and, with idle-sleep disabled, the holder
had no other reason to release. The wait blocks under mutex_tasks, so the whole
server stopped answering while /health still returned 200.
Express the request as kernel state instead. A waiter holds <arena>/want.lock
shared while it waits and drops it once it owns the token; the sleep decision
probes that lock non-blocking and releases the GPU while anyone is waiting. The
probe needs its own fd - flock treats two open file descriptions of one file
independently, so probing on the waiter's fd would convert our own lock rather
than conflict with it. Nothing can be missed, nothing goes stale, and a waiter
that dies is cleaned up by the kernel.
Probe from should_sleep() on the loop thread rather than from a warden thread.
Routing it through a flag is what made the first attempts fail: start_loop()
holds mutex_tasks from should_sleep() through the callbacks to the wait, so a
warden's request_yield() blocks on that mutex and is admitted only after the flag
has been consumed, latching a release for the next wake. Reading live state where
the decision is made has no edge to latch, and drops the warden's poll latency.
Two sleep-path bugs this exposed: wait_until_no_sleep() waited on !sleeping but
the loop clears req_stop_sleeping on the way in, so a loop that slept again
before the waiter ran stranded it forever - re-ask on every wake. And a task
queued after the waiter saw us awake could not wake us by itself, so sleep now
also breaks on a non-empty queue. Hold the GPU for 100 ms after a wake: the
request that woke us is not queued yet, and yielding at once only sends it round
again.
Measured on the RX 580 pod, two servers contending, 60 alternating handoffs:
0 stranded, median 0.309 s against the doorbell's 0.314 s.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PZz44SLQvTXMyWGio6t9DZ
Encodes the production config, fixed corpus slices, repeat/median discipline
and the noise floor, so the measurement method does not have to be
rediscovered each time. Runs the corpus prefill test through llama-server and
the llama-bench sweep as a controlled cross-check, with interleaved A/B.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PZz44SLQvTXMyWGio6t9DZ
The tiled-transpose fast path in ggml_vk_concat is a large prefill win but a
loss at decode. On a hybrid model like Qwen3.6-35B (gated delta-net on 30 of
40 layers) build_conv_state emits a transposed concat per recurrent layer, so
the fast path runs 30 times per decoded token.
At one token the transposed source is a single column, so there is no
uncoalesced stride left to fix, but the path still costs two dispatches and an
unconditional ggml_vk_sync_buffers - a full pipeline barrier that serializes
the command stream. Measured -17% tg256 on RX 580; the barrier, not the
half-empty transpose dispatch, is nearly all of it.
Require the transposed source to be at least one TILE_DIM wide. That source is
qkv_mixed transposed, so ne[0] is the ubatch token count: prefill keeps the
fast path, decode and small speculative drafts take the generic one.
pp8192 +11% retained, tg256 back to parity, generation byte-identical.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PZz44SLQvTXMyWGio6t9DZ
The vision encoder sits idle in VRAM on every text request. With
LLAMA_MMPROJ_ONDEMAND=1 the server keeps the encoder in RAM (released
after load) and brings it into VRAM just-in-time before mtmd_batch_encode,
releasing it again afterwards - so the encoder's VRAM is free for KV /
expert cache on the common text-only path.
When the encoder does not fit (that VRAM has been claimed), evict the LLM
backbone weights first: they are not needed while the encoder runs (the
decode that uses them happens afterwards), their host shadow is read-only
(cheap free, no D2H), and the KV / prompt cache is left untouched. Order
on the way back matters: free the encoder before re-uploading the weights
so the peak stays within VRAM.
Also wire the encoder release/restore into the VRAM arbiter: vram_go_cold
releases it; the just-in-time encode-path restore replaces the previous
restore in vram_ensure_warm.
Assisted-by: Claude
Add clip_release_device/clip_restore_device (and mtmd_release_device/
mtmd_restore_device wrappers over the vision + audio contexts) that free
the multimodal encoder's device weight buffer to a read-only host shadow
and rebuild it on demand, using the same shadow/free/reallocate pattern
as llama_model weights. No-op for a CPU-backed encoder. This lets the
server drop the ~hundreds-of-MiB vision encoder from VRAM when it is not
encoding an image.
Assisted-by: Claude
On device release with KV eviction, also call ggml_backend_free_scratch()
on each backend so a cold model drops its Vulkan compute preallocations
(reallocated lazily on the next compute via restore).
Also guard llama_context::memory_breakdown() against a freed scheduler:
evict_kv release resets sched to null, so a cold model that is then torn
down (e.g. terminated by the process manager) hit GGML_ASSERT(sched) in
ggml_backend_sched_get_buffer_type. Skip the compute-buffer accounting
when sched is null.
Assisted-by: Claude
Add an optional backend interface method free_scratch (with a public
ggml_backend_free_scratch wrapper) that frees transient/scratch device
memory a backend holds outside of any allocated buffer, keeping the
backend usable - the scratch is reallocated lazily on the next compute.
Implement it for the Vulkan backend (ggml_backend_vk_free_scratch): free
the prealloc_x/y/split_k/add_rms_partials and sync_staging device buffers
and reset their sizes, so an idle/cold model does not hold the vision or
matmul compute preallocations in VRAM. All other backends leave the hook
null (no-op).
Assisted-by: Claude
The recurrent (SSM/conv) state of hybrid models (e.g. Qwen3.5) was left
resident when a model's device buffers were released for on-demand VRAM
sharing - llama_memory_recurrent::release_device_buffers() was a no-op
default. Implement it (and restore_device_buffers) with the same
capture-host-shadow / free / reallocate pattern as llama_kv_cache, so
llama_memory_hybrid now evicts both its attention KV and its recurrent
state. The state is read-write, so its shadow is recaptured on every
release.
Assisted-by: Claude
Under LLAMA_SLEEP_EVICT_KV a cold model still held the scheduler's worst-case
compute buffer (hundreds of MiB, e.g. ~700 MiB for gemma-26B) plus the resident
experts. release_device(evict_kv) now also frees the sched (sched.reset() +
sched_need_reserve), and restore_device() rebuilds it via sched_reserve(), so a
fully-evicted model holds essentially no VRAM (gemma cold: 7147 -> 51 MiB).
Trades a heavier re-warm (weights+KV H2D + sched reserve) for the freed VRAM;
weights-only mode is unchanged (fast switch, keeps KV+compute).
Assisted-by: Claude
With LLAMA_SLEEP_EVICT_KV the KV cache device buffers are freed when a model
goes cold. update_slots() touches the KV before the decode-time vram_ensure_warm
(e.g. SWA models create a checkpoint that reads the KV via ggml_backend_tensor_get),
so the KV must already be resident by then. Move the restore to the sleep-wake
handler (handle_sleeping_state(false)), which runs before update_slots, fixing a
GGML_ASSERT(buffer) crash on iSWA models (gemma) when KV eviction is enabled.
Assisted-by: Claude
release_device_weights()/release_device_buffers() leave a null buffer in
ctxs_bufs while the device memory is released for on-demand VRAM sharing.
The memory-breakdown / total_size / clear paths iterated these and called
ggml_backend_buffer_get_size()/get_type()/clear() on the null buffer, tripping
GGML_ASSERT(buffer) and aborting on shutdown of a cold model. Skip null buffers
in llama_model::memory_breakdown and llama_kv_cache::{memory_breakdown,
total_size,clear}.
Assisted-by: Claude
Before uploading a model's weights, acquire the shared VRAM token (ring the
doorbell so any resident model releases first). Previously load uploaded weights
to VRAM before the arbiter was active, so loading a large model while another
(e.g. the warm 4B task model) held VRAM could exceed the budget and OOM.
- add vram_arena_open() (idempotent flock/doorbell setup) and
vram_acquire_for_load(), called from load_model() before
common_init_from_params().
- a coordinated load now stays warm holding the token and serves its first
request without a re-warm (drop the init-time go_cold cold-start).
Assisted-by: Claude
The request handler only calls wait_until_no_sleep() (which wakes a server out
of its sleeping state) when sleep_idle_seconds >= 0. But the VRAM arbiter's
cross-process doorbell can put a server to sleep even when idle-sleep is
disabled, so without this a doorbell-slept server would never wake and requests
to it would hang until timeout.
Do not bypass the wake path when LLAMA_SLEEP_VRAM_ONLY is set, so the arbiter no
longer depends on --sleep-idle-seconds being configured.
Assisted-by: Claude
Extend on-demand device residency to the KV cache so that when a model's KV
plus another model would not fit in VRAM, the KV can also be evicted to a host
shadow (D2H on release, H2D on restore) instead of only the weights.
- llama_memory_i: add release_device_buffers()/restore_device_buffers()
(default no-op). Implemented in llama_kv_cache (D2H shadow of the live
ctxs_bufs, freed and reallocated like the weights); llama_memory_hybrid and
llama_kv_cache_iswa delegate to their child caches.
- llama_context::release_device(evict_kv): also evict the memory's device
buffers when requested; restore_device() rebuilds them. Public API
llama_context_release_device gains an evict_kv flag.
- server: LLAMA_SLEEP_EVICT_KV=1 enables it. Off by default (weights-only),
since the KV shadow adds a D2H/H2D copy of the live cache each cycle.
Validated on RX 580 (Vulkan), 4B @ 32k ctx: weights-only cold VRAM 1750 MB
(KV stays); weights+KV cold VRAM 726 MB (KV freed, ~1 GB reclaimed). KV
survives the round-trip: prompt cache reused after the cycle (prompt_n 4 vs
42), correct output.
Assisted-by: Claude
Add release/restore of a model's GPU weight buffers (keeping a host shadow
and the KV cache) so several always-loaded llama-server processes can
time-share a single GPU without reloading or losing the prompt cache.
- llama-model: release_device_weights()/restore_device_weights() capture a
compact host shadow (stable iteration order, view-skipping) and free then
realloc the device weight buffers; weights_resident() query.
- llama-context: release_device()/restore_device() wrappers; decode() auto-
restores; public C API llama_context_release_device/restore_device.
- server: LLAMA_SLEEP_VRAM_ONLY makes idle-sleep release only the VRAM weights
(not a full unload/reload). A cross-process flock token in LLAMA_VRAM_ARENA
enforces "resident iff holds token"; an inotify doorbell forces the holder
to release on contention. The warden thread only touches the task queue, so
releases run on the loop thread and never race a decode.
Validated on RX 580 (Vulkan): two models share 8GB, never both resident,
correct output under contention, KV cache preserved (no re-prefill).
Assisted-by: Claude
The generic concat shader reads a transposed source (nb[1]==type_size) with an
uncoalesced stride, which is catastrophically slow on discrete GPUs (~5.6ms for
a 16MB concat on the RX 580 vs ~130us of memory bandwidth), a ~9% prefill
hotspot on Qwen3.5-4B (delta-net state concat).
Add a fast path for concat along dim 0 where one source is stored transposed and
the other source + dst are contiguous along dim 0: copy the contiguous source
with copy.comp and transpose the other source into the matching dst sub-region
with the existing tiled copy_transpose shader (shared-memory 32x32 transpose,
coalesced read+write). Reuses pipeline_cpy_* / pipeline_cpy_transpose_* with
custom push constants + doffset, no new shader. Falls back to the generic path
otherwise (gated on type/shape/contiguity + 16-bit doffset bound).
Assisted-by: opencode
Staging K/V through shared memory looks like an obvious win on GCN: without it
each rowgroup re-reads the whole K/V block from global memory, and with row_split
4 that pulls one 16KB block through a 16KB vector L1 four times.
Measured, it loses: -6.7% pp2048 at depth 16k and -7.4% at 32k on Polaris at head
size 128. The kvsh stride of D/4+1 dwords is 4 mod 32, which costs an 8-way LDS
bank conflict on wave64 - the +1 padding is tuned for warp32 - and the extra
shared memory eats occupancy this shader is already short of.
Comment only, no behaviour change. Leaving a note so the next person does not
spend a GPU on it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
mask_opt was disabled on AMD GCN for head sizes <= 256, but it is
beneficial there in high-context prefill: on fully-visible causal
blocks it skips the per-block mask load+add, and it skips fully-masked
blocks entirely. The attention op on GCN is compute-bound on the
softmax path (no matrix cores), so this cuts real work.
Verified lossless (perplexity bit-identical with it on vs off) and a
measured prefill win on Qwen3.5-35B-A3B (head_dim 256) on an RX 580:
pp2048 unchanged at short context, +8% @ 16k, +12% @ 32k, growing with
depth. Enable for GCN when HSK/HSV >= 256; the existing large-mask
conditions keep it off for decode.
Assisted-by: Claude
Offloaded expert weights are uploaded per split, and to decide which experts to
upload the copy callback reads the routing ids back to the host. The ids are
produced on the same device we are about to upload to, so the readback forces a
full pipeline flush - 47 of them per eval on a 48-layer MoE.
Once the batch draws enough experts the readback stops telling us anything. At
2048 tokens x 10 experts over 256 experts every expert comes back used, so the
used set is all true and the copy is a single whole-tensor range anyway.
Skip the readback when ggml_nelements(ids) >= 4 * n_expert and mark every expert
used. Uploading an expert that no id selects cannot change the result, since
mul_mat_id only reads the rows the ids point at, so this stays exact. With the
MoE cache on, the all-used set still goes through moe_cache->copy_experts, which
copies the cached experts device-to-device from the host slot maps and uploads
the rest - still exact, and no readback is needed for it. Decode is unaffected:
n_ids there is the number of experts per token, far below the threshold. The
small-batch MoE cache slot path has its own readback and is not touched.
The st.ids/st.used caching is unchanged: a repeat call with the same ids tensor
reuses the all-true set, which is what the threshold would decide again.
Port of dbc449afd ("sched: skip the MoE routing-ids readback when the batch
uses every expert"), whose target block in ggml_backend_sched_compute_splits was
removed upstream (#29943) in favour of this copy callback.
Measured (original, before the port) on an RX 580 (Vulkan, Polaris) running
Laguna-S-2.1 118B IQ2_M with all experts host-resident: pp2048 +4.7% at depth
32k, +0.9% at 16k, tg32 +3%.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PZz44SLQvTXMyWGio6t9DZ
The src1 power-of-two rescale in kernel_mul_mm_id is superseded by the
per expert-token amax scaling done in the compute graph
(ref: https://github.com/ggml-org/llama.cpp/pull/30861), which keeps the
MoE down-projection activations within f16 range before the op runs.
Drop the two-stage reduction kernels, their scratch buffer and the
GGML_PREC_F32 dispatch added in https://github.com/ggml-org/llama.cpp/pull/26223,
as well as the test cases that drive MUL_MAT_ID src1 out of f16 range.
Assisted-by: pi:llama.cpp/ggml-org/Qwen3.8-Flash-Next-GGUF (MacBook, M5 Max, 128GB)
* ui : move models manager to its own route, keep Discover as a dialog
Models now lives at #/models instead of a dialog. Discover opens as a
standalone dialog over it, and leaving the page returns to the chat it
was opened from.
Assisted by: Claude (Opus 5.5 Medium)
* ui : polish the models page layout and filters
Give the models page a page header with the Discover action, calmer
column labels, tinted section and family headers, and an empty state.
Load the model list on a direct visit, slide the detail pane over the
table below xl, and draw the filter toggles as one box matching the
context select, with room for their shadow.
Assisted by: Claude (Opus 5.5 Medium)
* ui : round the models table family headers
Assisted by: Claude (Opus 5.5 Medium)
* ui : fix models page navigation and Discover gating
Record the models page's way back after every navigation, so Back,
a typed URL or a reload no longer return to a stale chat. Escape
closes the detail pane, then leaves the page. Gate every Discover
entry point on one shared check, recheck the API key on /models,
skip the duplicate router fetch on a direct load, and share the
route load between pages.
Assisted by: Claude (Opus 5.5 Medium)
* ui : show model details in a drawer
Open the models manager's detail pane in a drawer, from the right on
desktop and from the bottom on a phone, instead of an inline side
panel and a full-screen cover. Pad the pane on both sides, drop its
close button on a phone, and assert that filtered rows are gone from
the page rather than merely hidden.
Assisted by: Claude (Opus 5.5 Medium)
* ui : make whole models table rows clickable
A click anywhere on a model or quant row opens it, unless it lands on
one of the row's own controls (load toggle, actions menu). The name
button stays the keyboard path to the same action.
Assisted by: Claude (Opus 5.5 Medium)
Use the dropdown menu surface (border, radius, background, shadow,
padding) for the hover card and the context gauge popup, and fix the
hover card open/close animation to use bits-ui data-state attributes.
Assisted-by: Claude (Opus 5.5 Medium)
The two T4 runners in the ggml-ci group (ggml-6-x86-vulkan-t4,
ggml-8-x86-nvidia-a10) share the same labels and can both pick up the
same jobs, so extend the per-runner results dir from #31046 to them.
Assisted-by: dsh:llama.cpp/ggml-org/Qwen3.8-Flash-Next-GGUF (MacBook, M5 Max, 128GB)
The generic scale pass in llama-model.cpp loads the per-expert
NVFP4 scale2 tensors (ffn_{up,gate,down}_exps_s) and the per-tensor
MXFP8 scales for the dense FFN and the shared expert, but
minimax-m3.cpp never forwarded any of them to the graph.
build_moe_ffn() defaults these to nullptr, so build_lora_mm_id()
skipped its w_s branch entirely and the routed-expert weights came
out global_scale times too large, swamping the residual (reported
G ~= 15123 for M3.1 Flash) and producing token soup.
Forward all scale tensors at the three call sites, matching the
pattern used by flint and the other NVFP4-capable MoE models.
Co-authored-by: origin-labs-ai <origin-labs-ai@users.noreply.github.com>
Add test-speculative-rejection-sampling to the ctest -E exclusion list of the macos-latest-arm64 CI job.
Assisted-by: dsh:llama.cpp/ggml-org/Flint-GGUF (RTX 5090 32GB, DDR5 128GB)
The amax POOL_1D reduction spans the whole matmul row, so a device cannot
compute it from an axis-0 split. The Meta split-state case reported
MIRRORED, which gives each device a partial max of its own slice. Drop
that case and deny -sm tensor for the arches which build the amax graph,
see [TAG_MOE_AMAX_SCALE].
Assisted-by: dsh:llama.cpp/ggml-org/Qwen3.8-Flash-Next-GGUF (MacBook, M5 Max, 128GB)
* ggml : allow MUL_MAT_ID -1 index to skip the computation
* nits
* update vulkan code
* ggml-webgpu: ensure all threads reach workgroup barriers in mul_mat_id_vec
(cherry picked from commit d8a4b9efb6e2b9abf3a866e847756b7e6bbc6b69)
* fix cuda
* openvino
* exclude CUDA from this PR
* add TODO tag
---------
Co-authored-by: yomaytk <yoshimura.masashi.frbs@gmail.com>
* tests : add speculative rejection draft sampling test
This commit adds a test for speculative rejection sampling test that
excercises the speculative rejection sampling mechanism in
common/speculative.cpp.
The motivation for this is have a test that verifies that the
speculative draft models emitts tokens that match the target models
distribution. So when rejection sampling occurs it should still
produce the same probability distribution as if we directly sampled
from the target model.
Running the test:
```console
$ ctest --test-dir build -R "^test-speculative-rejection-sampling$" \
--n_trials 5000 -V
```
This will generate two models, one target and one draft model.
For subsequent runs use the "fixture exclude any" option to avoid
recreating the models:
```console
$ ctest --test-dir build-cpu-debug -R \
"^test-speculative-rejection-sampling$" \
-FA "^speculative-models$" --n_trails 5000 -V
```
* remove draft model
* ci : fix data race between runners sharing the model cache
Two runners per mac mini share HOME, so concurrent jobs corrupt each
other downloads, converted models and venv under the shared MNT cache.
- gg_wget: download in a temp dir, publish with mv, so readers only
ever see complete files (-N cannot be combined with -O)
- convert, quantize and wikitext unzip now produce under a per-runner
scratch dir in OUT, published to MNT via hardlink when missing
- gguf-split and quantize tests get a per-runner TMP_DIR
- python venv moved from MNT to OUT
- metal, vulkan and webgpu mac jobs pass a per-runner OUT path
Assisted-by: dsh:llama.cpp/agent-qwen4 (RTX 5090 machine)
* ci : clean up scratch models after tests
Assisted-by: dsh:llama.cpp/agent-qwen4 (RTX 5090 machine)
* ci : fix q4_0 publish when a cached model exists
cp -n skips an existing file with exit code 1, so the ln fallback
aborted the qwen3 job under set -e. Publish through a unique temp
file with an atomic mv: readers never see a partial file and a
stale cached q4_0 gets replaced.
Assisted-by: dsh:llama.cpp/ggml-org/Qwen3.8-Flash-Next-GGUF (RTX 5090 machine)
* ci : cache-first gg_wget to avoid re-downloads
HF sends no Last-Modified, so wget -N re-downloads the full files on
every run. Skip the download when a non-empty cache file exists: the
HF CI artifacts are static. Downloads still go through a temp dir and
an atomic mv, so concurrent runners never read partial files.
Assisted-by: dsh:llama.cpp/ggml-org/Qwen3.8-Flash-Next-GGUF (RTX 5090 machine)
* HIP: add byte permute fallback when __builtin_amdgcn_perm is unavailable
amdgcnspirv rejects target-specific builtins unless they are guarded by `__builtin_amdgcn_is_invocable`, so 66 of 145 HIP files failed to build
Wrap the perm calls in a helper that keeps v_perm_b32 when the target has it and falls back to a portable byte select otherwise
* move perm helper to vendors/hip.h
* mtmd : lfm2 bilinear resize matches torchvision
Lfm2VlImageProcessor and d1-omni vision.py resize uint8 images with
torchvision (antialias=True). That is Pillow's algorithm, except that the
weights are int16 with as many fractional bits as the largest weight allows,
where Pillow uses 22 bits. The Pillow port was 1-2 levels off on about 1% of
the values of the d1-omni and LFM2.5-VL crops.
RESIZE_ALGO_BILINEAR_TORCH rounds the weights of the Pillow port to that
precision and scales them back to 22 bits, which gives the output of
torchvision with the same passes, and "bilinear" maps to it for LFM2 and
d1-omni. Other models keep the Pillow algorithms.
* mtmd : lfm2 bicubic resize matches torchvision
Lfm2VlImageProcessor resizes with torchvision for the bicubic checkpoints too
(d1-3B, LFM2-VL), so "bicubic" maps to RESIZE_ALGO_BICUBIC_TORCH for LFM2: the
Pillow bicubic filter (a = -0.5, as torchvision with antialias) with the int16
weights of torchvision. The Pillow port was 1-2 levels off on about 0.25% of
the values of the d1-3B crops.
* use bitfield, add hparams setter
---------
Co-authored-by: Xuan Son Nguyen <son@huggingface.co>
Add Soprano-80M and Soprano-1.1-80M (https://huggingface.co/ekwek/Soprano-1.1-80M)
as an audio generation model: a Qwen3 backbone plus a Vocos decoder that runs
as a gen-only mmproj.
- convert: detect Soprano checkpoints and export the backbone as qwen3 and the
decoder.pth Vocos weights as the mmproj
- vocab: add the soprano pre-tokenizer (lowercase, whitespace collapse,
per-digit split, [UNK] fallback)
- mtmd: add the Vocos graph and ISTFT, allow loading a gen-only mmproj for
Soprano and include the gen-audio context in mtmd_get_memory_usage
- mtmd-helper: add the Soprano pipeline; the decoder runs over the whole
utterance, so stream mode sends the audio in one chunk after flush()
- tts: document Soprano usage
Assisted-by: Codex
Assisted-by: Claude Opus 5.5
* hexagon : add SGN support
Assisted-by: OpenCode
* hexagon : optimize SGN with sign bit
Assisted-by: OpenCode
* hex-sgn: add f16
* hex-unary: update SGN to use generic unary_colapse and relax colapse checks to apply to more conditions
* hex-unary: improve dims colapsing and run it in the host-side solver
---------
Co-authored-by: Max Krasnyansky <maxk@qti.qualcomm.com>
The __builtin_amdgcn_wmma_f32_16x16x16_bf16_w32 (and _gfx12) builtins
take their A/B operands as vectors of short (_ExtVector<16, short> /
_ExtVector<8, short>), not __bf16. Declaring the fragments as
ext_vector_type(N) __bf16 fails to compile:
error: cannot initialize a parameter of type 'short
__attribute__((ext_vector_type(16)))' with an lvalue of type 'const
bf16x16_t' (vector of 16 '__bf16' values)
short and __bf16 are both 16-bit and the value is only ever passed to
the builtin via reinterpret_cast over raw bytes, so changing the element
type is bit-identical. The operand type has been short since the builtin
was introduced, so this works across LLVM/ROCm versions without a
version gate. Fixes HIP builds for RDNA3 (gfx1150/1151) and RDNA4
(gfx1200/1201).
Signed-off-by: Pablo Cholaky <waltercool@slash.cl>
* CUDA: fix build after chunked CUDA graphs (#30804)
* Apply suggestion from @JohannesGaessler
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
* Fix newlines in common.cuh
---------
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>